Semantic Representations in a Weightless Neural Network

AUTOR(ES)
DATA DE PUBLICAÇÃO

1999

RESUMO

The objective of this thesis is to present and evaluate a way of instilling semantic knowledge into a weightless neural model. The research developed uses neural state machines models (NSMM) to bind together the neural computational paradigm with knowledge representation paradigm. Starting from a critical analysis of existing models, it defines which generalisation capabilities are required by a neural state machine in order to emulate hierarchical relationships needed by semantic representations. Several generalisation algorithms, which combine the required generalisation capabilities and the application of newly defined rules of similarity, are presented and discussed. This effort has driven the definition of a novel spreading algorithm which applies different rules of similarities to node partitions responsible for identifying specific hierarchical relationships. This allowed queries can be posed to the neural state machine and be answered according to their semantics. The acquired generalisation capability is such that it does not require that all possible inter-relationships, which can be derived from the Knowledge Base itself, are worked out and trained in advance. The performance of the proposed neural state machine in generalisation tasks was experimentally tested and involved the creation of a new measurement for the quality of the results. Extensive testing and statistical analysis suggests that the proposed model is found to be robust with respect to variations both on the degree of connectivity and on the size of the training set.

ASSUNTO(S)

ciência cognitiva inter-relações em agrupamentos de padrões neural network redes neurais sem peso representação de conhecimento ciencia da computacao representação semântica distribuída e hierárquica algoritmos de generalização e de espalhamento

ACESSO AO ARTIGO

Documentos Relacionados